by Google · Gemini 3.5 family · best for production agent + coding backbone
Gemini 3.5 Flash is the first model in Google DeepMind's Gemini 3.5 family, launched GA at Google I/O on 2026-05-19 and positioned as the default backbone for agentic and coding workloads. Its headline trick: a Flash-tier model that beats last year's Pro tier and, on several agentic benchmarks Google publishes itself, beats the current Gemini 3.1 Pro — leading the field on MCP Atlas (83.6%) and posting Terminal-Bench 2.1 76.2% — while running roughly 4x faster than other frontier-class models at ~25% lower cost than 3.1 Pro. It keeps the full 1M context and native multimodal input. For a buyer: this is the production agent and coding engine on Google's stack, with 3.1 Pro reserved for the hardest pure-reasoning work.
| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 40.2% | deepmind.google 2026-05-19T00:00:00.000Z |
| MMMU | 83.6% | deepmind.google 2026-05-19T00:00:00.000Z |
| GPQA Diamond | 92.2% | buildfastwithai.com 2026-05-20T00:00:00.000Z |
| Terminal-Bench | 76.2% | deepmind.google 2026-05-19T00:00:00.000Z |
| MRCR Long Context | 77.3% | deepmind.google 2026-05-19T00:00:00.000Z |
| SWE-bench Verified | 80.8% | benchlm.ai 2026-05-24T00:00:00.000Z |
| Artificial Analysis Index | 55 | artificialanalysis.ai 2026-05-28T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“The model Google wants us to standardize on — fastest path to frontier-adjacent quality with full Vertex governance.”
For a CTO running an agent fleet, 3.5 Flash hits the sweet spot: 4x speed, ~25% under 3.1 Pro, best-in-class agentic benchmarks, and identical Vertex governance and Workspace integration to the Pro tier — no security trade-off. It's GA, so unlike 3.1 Pro there's no Pre-GA caveat. The strategic risk is the price drift from earlier Flash generations (3x Gemini 3 Flash Preview), which reshapes unit economics for teams scaling up from cheap Flash predecessors. Lock-in is limited to Google Cloud itself. Roadmap confidence is high with 3.5 Pro arriving.
“Google's positioning is sharp: Flash for agents, Pro for reasoning — and 3.5 Flash genuinely owns the agent slot.”
3.5 Flash is positioned as the production-agent default, and the benchmarks back the claim — leading MCP Atlas field-wide and beating its own Pro tier on agentic and coding tasks. Its competitive moat is the combination of speed, Search grounding, and Vertex distribution rather than any single eval lead. Against GPT-5 mini and Claude Sonnet it differentiates on agentic tooling and multimodal/chart reasoning. Market timing at I/O 2026 was strong, capturing the agent-stack conversation. The main strategic muddle is internal: 3.5 Flash overlapping 3.1 Pro on coding can confuse buyers on which to pick.
“Flat $1.50/$9 with a 90% cache discount beats Pro's tiered model — just don't pretend it's still cheap Flash.”
At $1.50/$9 with $0.15 cached input and batch at $0.75/$4.50, 3.5 Flash is a clear value step down from 3.1 Pro ($2/$12) — and crucially has no over-200K cliff, so cost is predictable on long prompts. For agent fleets, TCO genuinely beats running them on Pro. The honest caveat: it's ~3x the price of Gemini 3 Flash Preview and 6x Gemini 3.1 Flash-Lite, so teams migrating up from cheap Flash predecessors must re-model unit economics. Thinking tokens bill as output and can inflate cost on hard turns; cap the thinking budget where quality allows.
“Tool calls stay coherent across long loops and MCP just works — this is the most fun Gemini to build agents on.”
For a builder, 3.5 Flash is the best agentic experience in the family: stable tool state over long loops, real Model Context Protocol support, reliable terminal and browser automation, clean structured output via response schemas, and built-in code execution. Function-calling latency is noticeably faster than Pro, which materially improves agent UX. The 1M context skips a lot of RAG — but weak 1M recall means you still chunk for precision retrieval. SDK surface is identical to the rest of Gemini, so swapping it in is a one-line change. January 2025 cutoff forces fresh-data plumbing for current events.
“Snappy and capable for everyday work; it occasionally feels less thorough than 3.1 Pro on the hardest problems.”
In the Gemini app's default mode and on AI Pro, 3.5 Flash is the model most users actually touch. With thinking dialed down it's noticeably faster than 3.1 Pro Deep Think, and conversation quality is strong for everyday research, drafting, and analysis. On the hardest reasoning it can feel thinner than Pro. Multimodal input (charts, screenshots, video) is excellent. Refusal rate is similar to 3.1 Pro — stricter than OpenAI/Anthropic on some prompts. The 2026 UX overhaul (native macOS app, cleaner mobile) helped; Trustpilot remains mixed, mostly about caps and policy rather than this model.
“A 1M-token model that recalls 26.6% at 1M, with a ~61% hallucination rate — the context number is marketing, not memory.”
The agentic wins over 3.1 Pro are real and impressive, but the long-context story is oversold: MRCR collapses from 77.3% at 128K to 26.6% at 1M, so the headline window vastly exceeds reliable working memory. AA's ~61% Omniscience hallucination rate means Search grounding is doing heavy lifting on factual tasks. "Beats last year's Pro" is true but cherry-picked to agentic benchmarks; on pure reasoning (HLE, ARC-AGI-2) it clearly loses to 3.1 Pro. And the SWE-bench Verified ~80.8% figure is from aggregators, not Google's own card, which lists the harder SWE-Bench Pro at 55.1%. Good agent model — read the asterisks on context and factuality.
The full research notes behind this review — verified against primary sources.
Sparse mixture-of-experts (Gemini family), with parameter counts, expert counts, layers, and attention all undisclosed — honestly null. Disclosed and verifiable: a 1M-token context window, 65,536 max output tokens, a January 2025 knowledge cutoff, and native multimodal input (text, image, audio, video). It exposes a configurable thinking budget; the high AA TTFT (~18.9s) reflects thinking being on by default in that eval. Long-context behavior is the architecture's clearest limitation: MRCR v2 is 77.3% at 128K but collapses to 26.6% at the full 1M, so the window is larger than its reliable working memory.
Gemini 3.5 Flash is built for agents (cap_agentic 9.5, cap_function_calling 9.5): it leads the field on MCP Atlas (83.6%), posts Terminal-Bench 2.1 76.2%, OSWorld-Verified 78.4%, and Toolathlon 56.5%, holding tool state and routing across long loops better than any other Flash-tier model. Coding is strong (cap_coding 8.5): SWE-Bench Pro 55.1% on the official card, ~80.8% SWE-bench Verified per aggregators. Reasoning (8.5) and math (8.0) are good but trail 3.1 Pro (HLE 40.2% vs 44.4%, ARC-AGI-2 72.1% vs 77.1%). Vision and document/OCR lead (cap_vision 9.5, cap_document_ocr 9.0): MMMU-Pro 83.6% (#1) and CharXiv 84.2%. Long context (cap_long_context 8.5) has the 1M window but weak 1M recall (MRCR 26.6%). Real-time data (cap_realtime_data 9.0) via Google Search grounding. Creative writing (7.5) is workmanlike; instruction following (8.5) and multilingual (8.5) are solid; safety calibration (8.0) mirrors 3.1 Pro, with an AA-measured ~61% hallucination rate that makes grounding important for factual tasks.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| MCP Atlas | 83.6% | New high for Flash | #1 field-wide, beats 3.1 Pro (69.2%) | card |
| Terminal-Bench 2.1 | 76.2% | New high for Flash | Beats 3.1 Pro; 2nd behind GPT-5.5 | card |
| CharXiv | 84.2% | Significant gain | Field leader on chart reasoning | card |
| MMMU-Pro | 83.6% | Up from 2.5 Flash | #1 on llm-stats MMMU-Pro board | card |
| SWE-Bench Pro (Public) | 55.1% | New | Trails Opus 4.7 by ~9 pts | card |
| SWE-bench Verified | ~80.8% | Up ~10-15 pts vs 3 Flash | Ahead of Opus 4.6 per aggregator | BenchLM |
| OSWorld-Verified | 78.4% | New | Strong computer use | card |
| GDPval-AA | 1656 Elo | New | Beats 3.1 Pro | card |
| Humanity's Last Exam | 40.2% | New for Flash | Behind 3.1 Pro (44.4%), Opus 4.7 | card |
| ARC-AGI-2 | 72.1% | Strong | Behind 3.1 Pro (77.1%), GPT-5.5 (84.6%) | card |
| GPQA Diamond | ~92.2% | Up from 2.5 Flash | Trails 3.1 Pro (94.3%) | BFW |
| MRCR v2 (128K / 1M) | 77.3% / 26.6% | Trade-off vs Pro | 3.1 Pro wins long-context recall | card |
| Artificial Analysis Index | 55 | New | #8 of 148 | AA |
(MMLU-Pro, AIME 2025, LiveCodeBench, IFEval, SimpleQA, LMArena Elo are not on Google's card or not yet stable, left null.)
Sustained ~203.5 output tokens/sec (peaks above 280) makes 3.5 Flash roughly 4x faster than other frontier-class models — its core selling point for agent fleets and interactive UX. The ~18.9s AA TTFT looks high but reflects thinking enabled by default in that benchmark; with a low or zero thinking budget, first-token latency drops substantially and the model feels genuinely snappy. For high-throughput agent loops, browser automation, and streaming applications, the speed/quality ratio is the best in Google's lineup.
| Surface | Cost | Notes |
|---|---|---|
| API input | $1.50 / 1M tok | Flat, no 200K tier |
| API output | $9.00 / 1M tok | Thinking tokens billed as output |
| Cached input | $0.15 / 1M tok | 90% discount; + $1.00/1M tok/hour storage |
| Batch (in/out) | $0.75 / $4.50 | ~50% off; async |
| Search grounding | 5,000 prompts/mo free (shared across Gemini 3), then $14 / 1,000 queries | Live Google Search |
| Direct UI | $19.99/mo (AI Pro); $100 & $200/mo (Ultra) | |
| Free tier | "Free of charge" on AI Studio | RPD/RPM caps |
| Rate limits | Tiered | Raises with spend gates |
Proprietary, closed-weights. Available via the Gemini API (Google AI Studio), Vertex AI, Gemini CLI, Google Antigravity, and Vertex AI Agent Builder; resold through OpenRouter. Vertex AI provides VPC-SC, CMEK, audit logging, regional pinning (US, EU, Asia), and data residency. No open weights or self-hosting. Migration from 3.1 Pro or 2.x Flash is a model-name swap — the SDK surface is identical across the Gemini line. The pricing advantage (flat $1.50/$9, no 200K cliff) plus 4x speed makes it the natural default for production agents that don't need 3.1 Pro's reasoning ceiling.
Google Frontier Safety Framework with configurable filters. Paid API and Vertex inputs are not used to train models; the free AI Studio tier may be. Opt-out available. Compliance: SOC 2, HIPAA, GDPR, ISO 27001, FedRAMP, CCPA. Built-in content moderation. Refusal behavior mirrors 3.1 Pro — stricter than OpenAI/Anthropic on some categories. The AA-measured ~61% hallucination rate on the Omniscience eval is a real caution: pair factual workloads with Search grounding or retrieval.
SDKs in Python, TypeScript, Go, Java, Dart; integrations with LangChain, LlamaIndex, Vercel AI SDK, Genkit, Google ADK, and native Model Context Protocol. Powers the Gemini app default mode, Google AI Pro, Gemini CLI, Google Antigravity, and Vertex AI Agent Builder. Distribution across Google's surfaces plus its agent positioning make it mainstream and rising fast.
On agentic and coding/multimodal benchmarks Google publishes (MCP Atlas, Terminal-Bench, CharXiv, MMMU-Pro), yes. On pure reasoning (HLE, ARC-AGI-2) and long-context recall, no — 3.1 Pro wins. Pick by workload.
$1.50/$9 flat — about 3x Gemini 3 Flash Preview and 6x Gemini 3.1 Flash-Lite. The lift buys frontier-adjacent quality and 4x speed; re-model unit economics if migrating up from cheap Flash.
The window is 1M, but recall drops to 26.6% at 1M (77.3% at 128K). Treat ~128K-256K as the reliable working range and chunk beyond that.
AA measures ~61% on the Omniscience knowledge eval. Use Google Search grounding or retrieval for factual workloads.
~203 tok/s sustained (4x frontier peers). The high AA TTFT reflects thinking-on; lower the thinking budget for snappy first tokens.
No. Gemini is closed-weights, API/Vertex only.
No — it's a model-name swap; the SDK surface is identical across the Gemini line.
Wins on pure reasoning (HLE 44.4% vs 40.2%, ARC-AGI-2 77.1% vs 72.1%) and long-context recall; loses on speed, agentic tasks, and price. The intended internal split: Flash for agents, Pro for reasoning.
Stronger creative tone and edges SWE-Bench Pro; weaker on agentic tool benchmarks and lacks Search grounding. Sonnet for writing, 3.5 Flash for agents.
Comparable price and speed band; weaker native video and chart reasoning, no Workspace/Search-grounding tie-in. GPT-5.4 mini for OpenAI-stack agents; 3.5 Flash for Google-stack multimodal agents.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-05-27